Four Google Gemini 4 Argon AI agents defeated Minecraft’s Ender Dragon without human intervention after 27 hours of gameplay, marking what is believed to be a world first documented in a research paper released by Google researchers.
What the AI Agents Accomplished in Minecraft
- Four Gemini 4 Argon agents used the open-source Mineflayer library to coordinate and beat the Ender Dragon in just over 27 hours without pre-scouted seeds or human game logic.
- The autonomous run involved exchanging 229 in-game messages, executing 2,783 programs, and dying 605 times before securing the win.
- Google researchers tracked the session via a live dashboard, using a fifth agent to ensure rule-following while the primary team handled resource farming, exploration, and combat recovery on their own.
How the Autonomous Squad Navigated the Overworld
Reaching the final boss required completing a strict sequence of in-game trials without external human steering. To track down a Stronghold and craft Eyes of Ender, the agents had to harvest Ender Pearls and Ender Rods while working their way up from basic tools. Google placed the four-agent team into a random Minecraft seed with a singular goal and no pre-assigned strategy. The agents used the open-source Mineflayer library to control their player characters, relying entirely on in-game chat and shared memory to organize themselves.
Four Gemini 4 Argon agents beat Minecraft from scratch in 27 hours.Operating in a sandbox using only code and text, they wrote 2,783 programs, divided up jobs over chat, shared notes, and debugged errors together – all with no human help ↓https://t.co/uyIhDOQXGT
Roles emerged organically, with certain agents concentrating on building Enderman shelters to safely farm materials, while others secured food stocks or scouted adjacent terrain. When characters died, teammates shared combat tips and assisted in recovering lost gear. Over the course of the 27-hour session, the bots exchanged 229 messages, wrote and ran 2,783 separate programs, and accumulated 605 deaths. A fifth agent monitored the gameplay to ensure no cheating occurred.
The Computational Cost and Speed Comparisons
While the agents successfully completed the run, their execution times and efficiency fell well short of human speedrunners. The current Minecraft speedrun record for defeating the Ender Dragon stands at six minutes and 36 seconds, whereas the AI team required roughly 13.5 hours just to reach The End dimension, spending the remaining time battling the final boss. The computational resources required for the feat were substantial.
The agents consumed 1.35 billion input tokens during the playthrough. At standard consumer pricing, that volume would cost approximately $1,620, translating to roughly $60 per hour or $15 an hour for each of the four participating computer agents.
| Metric | AI Agent Playthrough | Human Speedrun Record |
|---|---|---|
| In-Game Deaths | 605 | N/A |
| Programs Executed | 2,783 | N/A |
| Token Consumption | 1.35 billion tokens | N/A |
Why Researchers Tested Agents Inside a Sandbox
The experiment served a broader purpose than simply testing video game proficiency. Google researchers stated they wanted to study how multi-agent teams coordinate within a transparent environment where every step remains visible. Observing these behaviors helps developers understand complex autonomous systems before deploying them into real-world scenarios. The core competencies required to survive and win in Minecraft—such as team coordination, failure recovery, and sustained goal-pursuit over a full day—mirror the exact skills autonomous agents need to operate safely alongside people.
